Two AI engineers of Couger ranked among the top 10 at CVPR 2019 workshop — a major international conference on computer vision
Two Couger Inc. (Head office: Shibuya-ku, Tokyo, CEO: Atsushi Ishii, hereinafter “Couger”) AI engineers won 3rd and 5th place in the “Pixel SkelNetOn” competition at CVPR2019（IEEE/CVF International Conference on Computer Vision and Pattern Recognition), one of the world’s premier conferences on computer vision.
Two engineers from the Couger Inc.’s AI team, Sabari Nathan and Priya Kansal, participated in the CVPR2019.
The two took part in “Pixel SkelNetOn”, a competition of AI models’ accuracy in recognizing skeletons from images.
For skeleton data extraction, data-generating algorithms called GANs are usually preferred. In contrast, two engineers from Couger Inc. have developed an original algorithm, called “Skeleton-Net”, by combining “U-net”, which is used for understanding an image at pixel level, and “HED”, which used for extracting image shapes with high accuracy, and succeeded in significant improving skeleton extraction accuracy (up from 62% to 77%).
Wide implications for computer vision technologies
The technology employed for skeleton identification has wide potential use cases in various fields. For example, when customers buy something at a convenience store where the shopping process usually involves the following steps: 1) stop in front of the products, 2) pick up the desired products, 3) put the products into the basket, 4) start walking away. For preventing the risk of not paying for products in an empty store — or in the cashier-less stores of the future — it’s important to be able to accurately detect if purchase happened or not. This can be determined by analyzing basic physical actions performed by a shopper’s limbs. Furthermore, skeleton identification can help identify if a person is a child or adult by height determination. This works not only for people but also for animal recognition of creatures like birds or giraffes. In other words, recognizing human body parts and body movement through skeleton recognition can play a very important role in many scenarios.
Before, installing motion capture was necessary to a quantify human body movements. But by using deep learning, it’s possible to recognize peoples’ actions without the need for special sensors in the camera.
“Non-verbal” interfaces to become very important for people in the future
Couger Inc. is developing a technology called Virtual Human Agent (VHA), which involved reproducing humans in a virtual space in the form of an AI agent that behaves like a human being.
AI agents such as VHAs represent the next generation of technology-human interfaces and are not just able to talk with proper expressions but analyze and react according to “non-verbal” signals.
For this purpose, Couger Inc. is developing advanced object recognition technology for human communication. The two AI researchers’ work and “Skeleton-Net”thesis, which was accepted at CVPR2019 are part of the on-going work. Moreover, it’s expected that AIs through learning to decipher nonverbal body language can better understand human movements and react appropriately in all situations.
About CVPR2019:
CVPR is an annual computer vision conference, where world-leading AI companies, including the likes of Facebook, Google, and universities, present papers on their on-going research. It brings together students, researchers and companies from people all over the world. CVPR2019 will be held from June 16th to 20th in California, USA.
